* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file tan_3510_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/tan/tan_3510_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"adv_api/math/tan.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TAN_TAN_C310_IMPL_H__
#endif
#ifndef IMPL_MATH_TAN_TAN_C310_IMPL_H
#define IMPL_MATH_TAN_TAN_C310_IMPL_H
#include "kernel_basic_intf.h"
#include "kernel_tensor.h"
#include "../../common/check.h"
namespace AscendC {
namespace TanInternal {
constexpr Reg::CastTrait TAN_CAST_TRAIT_F162F32 = {
Reg::RegLayout::ZERO, Reg::SatMode::UNKNOWN, Reg::MaskMergeMode::ZEROING, RoundMode::UNKNOWN};
constexpr Reg::CastTrait TAN_CAST_TRAIT_F322F16 = {
Reg::RegLayout::ZERO, Reg::SatMode::SAT, Reg::MaskMergeMode::ZEROING, RoundMode::CAST_RINT};
constexpr float PI_FOR_X_TODIV = 0.3183098733425140380859375;
constexpr float KPI_FIRS_PI_MULS = 0.0009670257568359375;
constexpr float PI_V2 = 3.140625;
constexpr float PI_DOWN = 1.57079637050628662109375;
constexpr float PI_DOWN_NEG = -1.57079637050628662109375;
constexpr float KPI_TWI_PI_MULS = 6.2771141529083251953125e-7;
constexpr float PI_RESDOWN_ADDS = 0.00000004371139000189375;
constexpr float PI_RESDOWN_ADDS_NEG = -0.00000004371139000189375;
constexpr float KPI_THIR_PI_MULS = 1.21644916362129151821136474609375e-10;
constexpr float KPI_FOR_PI_MULS = -1.0291767438275201129727065563201904296875e-13;
constexpr float TAN_RES_MULTI_SCA = 0.0698520831551998762793;
constexpr float TAN_RES_ADDICT_UP = -6.8711573651634203789;
constexpr float TAN_2ADDS = 61.20362572811089435388;
constexpr float TAN_3ADDS = -24.8048928861126769186219;
__simd_callee__ inline void TanRound(
Reg::RegTensor<float>& srcReg, Reg::RegTensor<float>& tmpReg, Reg::RegTensor<float>& roundReg,
Reg::RegTensor<float>& resReg, Reg::RegTensor<float>& downReg1, Reg::RegTensor<float>& downReg2, Reg::MaskReg mask)
{
k=round(x/π), x0=x-kπ, x0�?-π/2, π/2)
π=π_0+π_1+π_2+π_3+π_4 achieve final precision compensation.
Final solution�?
k = round(x * invpi)
x -= k * pi_0
x -= k * pi_1
down1 = x + pio2_high // pi/2 + x
down2 = x - pio2_high // x - pi/2
x -= k * pi_2
down1 -= k * pi_2
down2 -= k * pi_2
down1 -= down_adds
down2 += down_adds
x -= k * pi_3
down1 -= k * pi_3
down2 -= k * pi_3
x -= k * pi_4
down1 -= k * pi_4
down2 -= k * pi_4
*/
Reg::Muls(roundReg, srcReg, PI_FOR_X_TODIV, mask);
Reg::Truncate<float, RoundMode::CAST_RINT, Reg::MaskMergeMode::ZEROING>(roundReg, roundReg, mask);
Reg::Muls(tmpReg, roundReg, PI_V2, mask);
Reg::Sub(resReg, srcReg, tmpReg, mask);
Reg::Muls(tmpReg, roundReg, KPI_FIRS_PI_MULS, mask);
Reg::Sub(resReg, resReg, tmpReg, mask);
Reg::Adds(downReg1, resReg, PI_DOWN, mask);
Reg::Adds(downReg2, resReg, PI_DOWN_NEG, mask);
Reg::Muls(tmpReg, roundReg, KPI_TWI_PI_MULS, mask);
Reg::Sub(resReg, resReg, tmpReg, mask);
Reg::Sub(downReg1, downReg1, tmpReg, mask);
Reg::Sub(downReg2, downReg2, tmpReg, mask);
Reg::Adds(downReg1, downReg1, PI_RESDOWN_ADDS_NEG, mask);
Reg::Adds(downReg2, downReg2, PI_RESDOWN_ADDS, mask);
Reg::Muls(tmpReg, roundReg, KPI_THIR_PI_MULS, mask);
Reg::Sub(resReg, resReg, tmpReg, mask);
Reg::Sub(downReg1, downReg1, tmpReg, mask);
Reg::Sub(downReg2, downReg2, tmpReg, mask);
Reg::Muls(tmpReg, roundReg, KPI_FOR_PI_MULS, mask);
Reg::Sub(resReg, resReg, tmpReg, mask);
Reg::Sub(downReg1, downReg1, tmpReg, mask);
Reg::Sub(downReg2, downReg2, tmpReg, mask);
}
__simd_callee__ inline void TanPolynomialApproximation(
Reg::RegTensor<float>& dstReg, Reg::RegTensor<float>& tmpReg, Reg::RegTensor<float>& roundReg,
Reg::RegTensor<float>& resReg, Reg::RegTensor<float>& downReg1, Reg::RegTensor<float>& downReg2, Reg::MaskReg mask)
{
tan(x) = xP(x) / ((π/2 - x)(π/2 + x)Q(x))
P(x) = (x^2 * R0 + R1) * x^2 + R2
Q(x) = x^2 * R3
R0 = 0.0698520831551998762793
R1 = -6.8711573651634203789
R2 = 61.20362572811089435388
R3 = -24.8048928861126769186219
*/
Reg::Mul(roundReg, resReg, resReg, mask);
Reg::Muls(tmpReg, roundReg, TAN_RES_MULTI_SCA, mask);
Reg::Adds(tmpReg, tmpReg, TAN_RES_ADDICT_UP, mask);
Reg::Mul(tmpReg, tmpReg, roundReg, mask);
Reg::Adds(tmpReg, tmpReg, TAN_2ADDS, mask);
Reg::Mul(tmpReg, tmpReg, resReg, mask);
Reg::Adds(roundReg, roundReg, TAN_3ADDS, mask);
Reg::Mul(roundReg, roundReg, downReg1, mask);
Reg::Mul(roundReg, roundReg, downReg2, mask);
Reg::Div(dstReg, tmpReg, roundReg, mask);
}
template <typename T>
__simd_vf__ inline void TanCompute(__ubuf__ T* dstUb, __ubuf__ T* srcUb, uint32_t sreg, uint16_t repeatTimes)
{
constexpr uint32_t stride = GetVecLen() / sizeof(float);
Reg::MaskReg mask;
Reg::RegTensor<T> srcReg;
Reg::RegTensor<float> castReg;
Reg::RegTensor<float> roundReg;
Reg::RegTensor<float> tmpReg;
Reg::RegTensor<float> downReg1;
Reg::RegTensor<float> downReg2;
Reg::RegTensor<float> resReg;
Reg::RegTensor<float> dstReg;
for (uint16_t i = 0; i < repeatTimes; ++i) {
mask = Reg::UpdateMask<float>(sreg);
if constexpr (sizeof(T) == sizeof(half)) {
Reg::LoadAlign<T, Reg::LoadDist::DIST_UNPACK_B16>(srcReg, srcUb + i * stride);
Reg::Cast<float, T, TAN_CAST_TRAIT_F162F32>(castReg, srcReg, mask);
} else {
Reg::LoadAlign(castReg, srcUb + i * stride);
}
TanRound(castReg, tmpReg, roundReg, resReg, downReg1, downReg2, mask);
TanPolynomialApproximation(dstReg, tmpReg, roundReg, resReg, downReg1, downReg2, mask);
if constexpr (sizeof(T) == sizeof(half)) {
Reg::Cast<T, float, TAN_CAST_TRAIT_F322F16>(srcReg, dstReg, mask);
Reg::StoreAlign<T, Reg::StoreDist::DIST_PACK_B32>(dstUb + i * stride, srcReg, mask);
} else {
Reg::StoreAlign(dstUb + i * stride, dstReg, mask);
}
}
}
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void TanImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t calCount)
{
if ASCEND_IS_AIC {
return;
}
CheckTensorPosition(dstTensor, "dstTensor", "VECIN, VECOUT, VECCALC");
CheckTensorPosition(srcTensor, "srcTensor", "VECIN, VECOUT, VECCALC");
CheckCalCount(calCount, "calCount", srcTensor, "srcTensor", "Tan");
CheckCalCount(calCount, "calCount", dstTensor, "dstTensor", "Tan");
static_assert(SupportType<T, half, float>(), "current data type is not supported on current device!");
constexpr uint32_t stride = GetVecLen() / sizeof(float);
uint16_t repeatTimes = CeilDivision(calCount, stride);
__ubuf__ T* dstUb = (__ubuf__ T*)dstTensor.GetPhyAddr();
__ubuf__ T* srcUb = (__ubuf__ T*)srcTensor.GetPhyAddr();
TanInternal::TanCompute<T>(dstUb, srcUb, calCount, repeatTimes);
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void TanImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const uint32_t calCount)
{
CheckTensorPosition(sharedTmpBuffer, "sharedTmpBuffer", "VECIN, VECOUT, VECCALC");
TanImpl(dstTensor, srcTensor, calCount);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TAN_TAN_C310_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TAN_TAN_C310_IMPL_H__
#endif